Noisy smoothing image source identification

Abstract

Feature based image source identification plays an important role in the toolbox for forensics investigations on images. Conventional feature based identification schemes suffer from the problem of noise, that is, the training dataset contains noisy samples. To address this problem, we propose a new Noisy Smoothing Image Source Identification (NS-ISI) method. NS-ISI address the noise problem in two steps. In step 1, we employ a classifier ensemble approach for noise level evaluation for each training sample. The noise level indicates the probability of being noisy. In step 2, a noise sensitive sampling method is employed to sample training samples from original training set according to the noise level, producing a new training dataset. The experiments carried out on the Dresden image collection confirms the effectiveness of the proposed NS-ISI. When the noisy samples present, the identification accuracy of NS-ISI is significantly better than traditional methods.